GPNet: Infrared Small Target Detection via Global Information Enhancement and Position Attention Guidance

Yuye Zhang, Xiuhong Li, Ying Zheng, Boyuan Li, Kangwei Liu, Jian Ma, Dangxuan Wu

Abstract: Infrared small target detection (IRSTD) is widely used in practical applications such as military early warning and coastal defense monitoring. How to accurately and automatically detect small targets in complex scenes is still challenging. Some convolutional neural network (CNN) methods based on U-shaped structure focus on the extraction of local small target features and ignore the attention and processing of background information, resulting in the network lacking effective perception of global information. At the same time, the continuous down-sampling of feature maps by the network in the encoding stage can easily cause the loss of small target feature information. Therefore, the network is prone to miss detection or false alarm of small targets in complex backgrounds. To solve the above problems, we propose a global information enhancement and position attention guidance network (GPNet). Specifically, it is implemented based on the encoder-decoder architecture. we design a global information enhancement (GIE) module in the encoding stage, aiming to fully capture and dynamically fuse global information. In the decoding stage, we designed the position attention guidance (PAG) module to use position prior information to compensate for the small target information lost during the downsampling process and guide the network to quickly locate the real target area. A large number of experiments have proven that the performance of GPNet has been significantly improved compared to other advanced IRSTD networks. It is worth noting that the target detection rate of GPNet on the NUAA-SIRST public dataset can reach 100%.